U.S. patent application number 10/943551 was filed with the patent office on 2005-08-04 for scalable mpeg video/macro block rate control.
This patent application is currently assigned to SONY CORPORATION. Invention is credited to Lee, Hung-Ju.
Application Number | 20050169369 10/943551 |
Document ID | / |
Family ID | 34812093 |
Filed Date | 2005-08-04 |
United States Patent
Application |
20050169369 |
Kind Code |
A1 |
Lee, Hung-Ju |
August 4, 2005 |
Scalable MPEG video/macro block rate control
Abstract
A scaleable macro block rate control method particularly
well-suited for MPEG video. There is provided a method to easily
derive a quantization parameter (QP) value using information such
as bit usage, previous QP values and SAD values from the past
encoded and future frames. The method utilizes quantization
estimation techniques based on statistical relationships between
different intensity measures, such as distortion intensity,
absolute difference intensity and mean of absolute difference
intensity. The method is well-suited to applications utilizing MPEG
video such as MPEG-1, MPEG-2, MPEG-4, JVT/H.264 standards and so
forth.
Inventors: |
Lee, Hung-Ju; (Pleasanton,
CA) |
Correspondence
Address: |
O'BANION & RITCHEY LLP/ SONY ELECTRONICS, INC.
400 CAPITOL MALL
SUITE 1550
SACRAMENTO
CA
95814
US
|
Assignee: |
SONY CORPORATION
SONY ELECTRONICS INC.
|
Family ID: |
34812093 |
Appl. No.: |
10/943551 |
Filed: |
September 17, 2004 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
60554533 |
Mar 18, 2004 |
|
|
|
60541340 |
Feb 3, 2004 |
|
|
|
Current U.S.
Class: |
375/240.03 ;
375/240.24; 375/E7.139; 375/E7.162 |
Current CPC
Class: |
H04N 19/124 20141101;
H04N 19/14 20141101 |
Class at
Publication: |
375/240.03 ;
375/240.24 |
International
Class: |
H04N 007/12 |
Claims
What is claimed is:
1. A method for scalable rate control for MPEG video, comprising:
calculating an average bit count using a plurality of future
frames; calculating a previous bit count using a past encoded
frame; calculating a target Qscale using the average bit count and
previous bit count to adjust a deficit/surplus bit rate budget;
refining Qscale using SAD statistics from past encoded frames and
current buffer status; and bounding Qscale in the range of
pre-defined MinQP and MaxQP values.
2. A method as recited in claim 1, wherein said MinQP value is
selected for maintaining maximum video quality for encoding a video
sequence.
3. A method as recited in claim 1, wherein said Qscale value is a
function of bit rate and frame rate.
4. A method as recited in claim 1, further comprising: adjusting a
bit rate budget; estimating a target bit rate for a current frame;
computing a Qscale value for first and subsequent I-VOPs; and
computing a Qscale value for P-VOPs.
5. A method as recited in claim 4, wherein said step of adjusting a
bit rate budget comprises: allocating bits as a function of SAD
values of P-VOP and MAD values of I-VOP by calculating total
available bit budget in a RC window and bit budget for a frame
before encoding a frame.
6. A method as recited in claim 4, wherein target bit rate
estimation comprises: determining bit allocation for a frame by
frame type and the associated SAD/MAD.
7. A method as recited in claim 6, wherein for each picture type,
all of its associated SAD/MAD values are summed, and the bit rate
budget for a frame with P-type is proportional to its MAD with
respect to overall MAD within a RC sliding window.
8. A method as recited in claim 1, further comprising encoding a
video frame or object using QP obtained in a pre-encoding step if
either frame- or object-level rate control can be activated.
9. A method for scalable macro block rate control, comprising:
performing an initialization step; performing a boot up step;
performing a mode decision step; performing a complexity step;
calculating a QP; performing a coding step; performing a update
step; and performing an after update step.
10. A method as recited in claim 9, wherein said initialization
step comprises: initializing data members; setting image dimension
information; and allocating working memory.
11. A method as recited in claim 9, wherein said boot up step
comprises: obtaining a picture level target QP from a picture level
rate control; calculating, for each macro block, content complexity
and average MAD; determining an intensity base; and determining a
local adjust range.
12. A method as recited in claim 9, wherein said step of mode
decision comprises: calculating InterAD and IntraAD values;
obtaining IntraAD values from a mode decision module; calculating
macro block level QP in I-slice; and sending the macro block level
QP to the mode decision module.
13. A method as recited in claim 9, wherein complexity step
comprises: obtaining IntraAd and InterAd from mode decision module;
calculating MB-level Qps, IntraQp and InterQp, for Inter model and
Intra mode, respectively in P-slice and B-slice; sending said Qp
values to the mode decision module.
14. A method as recited in claim 9, wherein said coding step
comprises encoding a current macro block.
15. A method as recited in claim 9, wherein said update step
comprises: accumulating Qp, Ad and number of coded macroblocks.
16. A method as recited in claim 9, wherein said after update step
comprises: calculating distortion of a current frame; calculating
average Ad for a next frame; and calculating average LocalAdj for
the next frame.
17. A method of deriving a quantization parameter for macroblocks
and pictures, comprising: calculating statistical difference and
distortion based on intensity; and generating a quantization
parameter in response to a comparison of said difference and
distortion statistics.
18. A method as recited in claim 17, wherein said calculating
generates values for distortion intensity, absolute difference
intensity and mean absolute difference intensity.
19. A method as recited in claim 17, wherein an intensity base is
added to the calculated values generated.
20. A method as recited in claim 19, wherein said intensity base
comprises a predetermined portion of an average mean absolute
difference intensity.
Description
CROSS-REFERENCE TO RELATED APPLICATIONS
[0001] This application claims priority from U.S. provisional
application Ser. No. 60/541,340 filed on Feb. 3, 2004, incorporated
herein by reference in its entirety and from U.S. provisional
application Ser. No. 60/554,533 filed on Mar. 18, 2004,
incorporated herein by reference in its entirety.
STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH OR DEVELOPMENT
[0002] Not Applicable
INCORPORATION-BY-REFERENCE OF MATERIAL SUBMITTED ON A COMPACT
DISC
[0003] Not Applicable
NOTICE OF MATERIAL SUBJECT TO COPYRIGHT PROTECTION
[0004] A portion of the material in this patent document is subject
to copyright protection under the copyright laws of the United
States and of other countries. The owner of the copyright rights
has no objection to the facsimile reproduction by anyone of the
patent document or the patent disclosure, as it appears in the
United States Patent and Trademark Office publicly available file
or records, but otherwise reserves all copyright rights whatsoever.
The copyright owner does not hereby waive any of its rights to have
this patent document maintained in secrecy, including without
limitation its rights pursuant to 37 C.F.R. .sctn. 1.14.
BACKGROUND OF THE INVENTION
[0005] 1. Field of the Invention
[0006] This invention pertains generally to video coding and
decoding techniques, and more particularly to methods for scalable
macro block layer rate control and picture layer rate control for
MPEG video.
[0007] 2. Description of Related Art
[0008] In real video applications, where either a real-time
streaming mode or a non-real-time batch mode is used, constant
video quality is a goal in developing a rate control scheme.
Although the first order and the second order rate distortion
models for these two modes provide a good foundation for target bit
rate estimation and quantization parameter (QP) estimation, the
fluctuation in target bit rate and quantization parameter values
derived from the model generates unstable video quality, resulting
in worse viewing quality.
[0009] Therefore, a need exists for a scalable rate control method
that is simple to implement and that results in better viewing
quality. The present invention provides these benefits and
overcomes the drawbacks of prior methods.
BRIEF SUMMARY OF THE INVENTION
[0010] Video coding and decoding techniques are described for
scalable macro block (MB) layer rate control and picture layer rate
control for MPEG video, including but not limited to, MPEG-1,
MPEG-2, MPEG-4 and JVT/H.264 standards. According to an aspect of
the invention, there is provided a method to easily derive a
quantization parameter (QP) estimation value using information such
as bit usage, previous QP values and sum of absolute difference
(SAD) values from both past encoded and future frames. In one
embodiment, the method comprises the steps of calculating the
average bit count (AvgTargetBitP) obtained in a "bit allocation"
module (using future frames); calculating the previous bit count
(ActualBitPrv) (using past encoded frame); calculating the target
quantizer scale (Qscale) using AvgTargetBit and ActualBitPrv to
adjust the deficit/surplus bit rate budget; refining Qscale using
SAD statistics from past encoded frames and current buffer status;
and bounding Qscale in the range of pre-defined MinQP and MaxQP
values.
[0011] Another aspect of the invention is to provide for scalable
macroblock rate control for quality improvement based on picture
contents and coding complexity. In one embodiment, the method
comprises the steps of initialization wherein, for example, data
members can be initialized, image dimension information can be set,
and working memory can be allocated; performing an
iRateCtrlMbBootUp step which can comprise obtaining the picture
level target QP from picture level rate control, calculating each
MB content complexity (MbMad) and average Mad, determining the
intensity base, and determining the local adjust range; performing
a Mode Decision step which comprises calculating InterAD and
IntraAD, and possibly determining the final mode and its QP;
performing a iRateCtrlMbCalcComplexitylonly step which comprises
obtaining IntraAd from mode decision module, calculating the
MB-level QP in I-slice, sending this QP back to mode decision
module; performing a iRateCtrlMbCalcQp step which comprises
obtaining IntraAd and InterAd from the mode decision module,
calculating two MB-level QPs, IntraQp and InterQp, for Inter mode
and Intra mode, respectively in P-slice, and sending the two QPs to
the mode decision module; performing an iecMbCoding step which
comprises encoding the current MB; performing a vRateCtrlMbUpdate
step which comprises accumulating QP, Ad and the number of coded
MBs, checking mode coding where if there is no mode MB coding,
returning to the mode decision module or otherwise proceeding to
the next step; and performing a vRateCtrlMbAfter step which
comprises calculating the distortion of the current frame,
calculating the average Ad for the next frame, and calculate the
average LocalAdj for the next frame.
[0012] Further aspects of the invention will be brought out in the
following portions of the specification, wherein the detailed
description is for the purpose of fully disclosing preferred
embodiments of the invention without placing limitations
thereon.
BRIEF DESCRIPTION OF THE SEVERAL VIEWS OF THE DRAWING(S)
[0013] The invention will be more fully understood by reference to
the following drawings which are for illustrative purposes
only:
[0014] FIG. 1 is flow diagram of an embodiment of a method for
scalable rate control of MPEG video according to an embodiment of
the present invention.
[0015] FIG. 2 is a graph showing QP as a function of buffer
deviation corresponding to the left half of Table 2.
[0016] FIG. 3 is a graph showing QP as a function of buffer
deviation corresponding to the right half of Table 2.
[0017] FIG. 4 is a graph showing QP as a function of buffer
deviation corresponding to the left half of Table 3.
[0018] FIG. 5 is a graph showing QP as a function of buffer
deviation corresponding to the right half of Table 3.
[0019] FIG. 6 is a flow diagram of an embodiment of a method for
scalable macro block control according to an embodiment of the
present invention.
[0020] FIG. 7 is a graph showing the dynamic range distribution of
QP as a function of picture syntax QP.
DETAILED DESCRIPTION OF THE INVENTION
[0021] Referring more specifically to the drawings, for
illustrative purposes the present invention is embodied in the
described methods and techniques. It will be appreciated that the
invention may vary as to the specific steps and sequence, without
departing from the basic concepts as disclosed herein.
[0022] A. Scalable Rate Control for MPEG Video
[0023] The following discussion illustrates an example embodiment
of the present invention. It will be appreciated that labels,
terms, sequences and other specific parameters are used in
connection with the example described herein and may be changed
without departing from the scope of the invention.
[0024] 1. Bit Rate Control Methodology
[0025] One aspect of the invention pertains to easily deriving a QP
value using information such as bit usage, previous QP values and
SAD values from the past encoded and future frames. Referring to
FIG. 1, in one embodiment, this is accomplished by carrying out the
following steps:
[0026] At step 10, the average bit count (AvgTargetBitP) obtained
in a "bit allocation" module is calculated using future frames.
Next, at step 12, the previous bit count (ActualBitPrv) is
calculated using a past encoded frame. At step 14, the target
Qscale is calculated using AvgTargetBit and ActualBitPrv to adjust
the deficit/surplus bit rate budget. Next, at step 16, Qscale is
refined using SAD statistics from past encoded frames and current
buffer status. Finally, at step 18, Qscale is bounded in the range
of pre-defined MinQp and MaxQp. Elements of this process are
described below.
[0027] 2. Execution Phase 0: Initialization
[0028] In this phase, initialization takes place, including the
initial buffer status, buffer convergence factor, data arrays for
collecting statistics, and specifying the minimum Qscale,
RateControl sliding window.
[0029] 2.1 Minimum Qscale Value Setting
[0030] The Minimum QP setting is for maintaining maximum video
quality for encoding a video sequence. The Qscale value is
dependent on the bit rate and frame rate. In the invention, bits
per MB are used as a threshold to determine the minimum Qscale, as
shown in Table 1.
[0031] 3. Execution Phase 1: Pre-Encoding
[0032] In this phase, the encoder performs the following
functions:
[0033] (a) Bit rate budget adjustment;
[0034] (b) Target bit rate estimation for the current frame;
[0035] (c) Qscale calculation for the first I-VOP and subsequent
I-VOPs; and
[0036] (d) Qscale calculation for P-VOPs.
[0037] 3.1 Bit Rate Budget Adjustment
[0038] Bit allocation is performed based on the SAD values of
P-VOP, B-VOP and MAD values of I-VOP. In bit allocation the first
step is to calculate the total available bit budget in a Rate
Control (RC) window and bit budget for a frame before encoding a
frame. To monitor these two values, a sliding window based bit
allocation is used, for example as follows:
[0039] // total number of bits available for this RC sliding
window
TotalBudget=(DefaultGovSec*Bit rate*RC_SPAN+TotalLoan),
[0040] Wherein TotalLoan=(CurrentOccupancy-InitialOccupancy) is the
amount of under-spending (TotalLoan>0) or overspending the bit
budget (TotalLoan<0).
[0041] 3.2 Target Bit Rate Estimation
[0042] Next, the bit allocation for a frame is determined by its
frame type and the associated SAD/MAD. First, for each picture
type, all of its associated SAD/MAD values are summed, and
basically the bit budget for a frame with P-type or B-type is
proportional to its MAD with respect to the overall MAD within a RC
sliding window. To maintain a minimum quality, an average bit
budget for a frame is calculated, and the MAD variation of a frame
contributed to the bit budget is bounded, such as by minus and plus
15% as follows:
AvgMAD=TotalMAD/Num_of_P_Vops
AvgTargetBitP=(TotalBudget/((m_dRCWindowSize+3.7)*RC_SPAN))
Variation=limit(-0.15, (TargetMADt-AvgMAD)/AvgMAD, 0.15)
TargetBits=(int)(AvgTargetBitP*(1.0+dVariation))
[0043] After this bit allocation, the buffer fullness is taken into
account by adjusting the target bits toward the initial buffer
occupancy as follows:
Cushion=(PseudoBufferSizeTop-Occupancy)
Fullness=(Occupancy-PseudoBufferSizeBottom)
TargetBits=(TargetBitCur*(Cushion+ConvergeFactor*Fullness)/(ConvergeFactor-
*Cushion+Fullness))
[0044] 3.3 QP Value for the First I-VOP
[0045] In a real application, the video quality of the first VOP
plays an important role in determining the first impression of a
user viewing a decompressed video sequence. Unlike MPEG-4 committee
code which requires input for the first I, P and B VOPs, according
to the rate control scheme of the present invention, a simple but
effective assumption is made to provide a reasonable video quality
of the first frame without user interaction. The QP value is
determined based on the following assumption: For example, if the
DefaultCompressionRatio is equal to 10:1, then DefaultQScale is
reasonably equal to 5 so that the Qscale for the first I-VOP is
calculated as follows:
DefaultTargetBits=Height*Width*8/DefaultCompressionRatio; // for
Luminance part
DefaultTargetBits+=(Height*Width/2*8/(DefaultCompressionRatio*2);
// for Chroma part
TargetBits=min(TargetBits, (VbvBufSize-InitialOccupancy)*0.2)
Qscale=DefaultTargetBits*DefaultQScale/TargetBits
[0046] To further refine the calculation, two difference encoding
modes (streaming mode and non-streaming mode) are preferably
utilized to lower the Qscale further, and bounded, such as by
Qscale=8 for minimum image quality in this example.
[0047] 3.4 QP Value for the Subsequent I-VOP
[0048] Besides the first I-VOP, QP values for the subsequent I-VOPs
also need to be considered. Although an RD model for I-type VOP can
be developed, due to its high complexity and without considering
video quality of its neighboring frames, the present invention
employs a simple scheme to determine the QP value of I-VOP. In
accordance with the invention, the QP value depends on three
factors: (1) previous QP value, (2) calculated QP value (target QP)
obtained in Section 3.2, and (3) the current buffer fullness as
follows:
1 If (Buffer occupancy - Initial Occupancy) > 0, implying
under-spending bits, then If (PrevQP > TargetQP) TargetQP =
TargetQP * ( 1 - (1-deviation){circumflex over ( )}2) + PrevQP *
(1-deviation){circumflex over ( )}2) and is bounded by (TargetQP +
PrevQP ) /2; If (PrevQP <= TargetQP) TargetQP = PrevQP * ( 1 -
(1-deviation){circumflex over ( )}2) + TargetQP *
(1-deviation){circumflex over ( )}2) where deviation = (Occupancy -
InitialOccupancy) / (VbvBufSize-InitialOccupancy).
[0049] In Table 2 and FIG. 2 and FIG. 3, two examples are shown to
demonstrate the Qscale calculations under these conditions. The
left hand side shows that if PrevQP (e.g., 28)>TargetQP (e.g.,
8), then the final QP is shown in column 5. The right hand side of
the table, on the other hand, shows that if PrevQP (e.g.,
3)<TargetQP (e.g., 8), and the final QP is calculated in the
last column.
[0050] On the other hand,
2 if (Buffer occupancy - Initial Occupancy) < 0 // implying
over-spending bits If (PrevQP > TargetQP) TargetQP = PrevQP *
(deviation{circumflex over ( )}2) + TargetQP *
(1-deviation{circumflex over ( )}2) If (PrevQP <= TargetQP)
TargetQP = TargetQP * (deviation{circumflex over ( )}2) + PrevQP *
(1-deviation{circumflex over ( )}2) where deviation =
(InitialOccupancy - Occupancy) / (InitialOccupancy).
[0051] In Table 3 and FIG. 4 and FIG. 5, two examples are shown to
demonstrate the Qscale calculations under these conditions. The
left hand side shows that if PrevQP (e.g., 28)>TargetQP (e.g.,
8), then the final QP is shown in column 4. The right hand side of
the table, on the other hand, shows that if PrevQP (e.g.,
3)<TargetQP (e.g., 8), then the final QP is calculated in the
last column.
[0052] 3.5 QP Value for the First P-VOP
[0053] The QP value for the first P-VOP is calculated as follows.
First, the target bit rate for this VOP is obtained by the bit
allocation module. Then its QP can be derived from the following
equation empirically. To prevent the occurrence of a sudden quality
change, this QP is bounded by its previous I-VOP's Qscale, and
further bounded by MinQP and MaxQP.
dQCur=(m_iActualBitPrvl*m_iQPrvl)/(8.0*m_iTargetBitCur)
dQCur=limit((double)m_iQPrvl, m_dQCur, (double)m_iQPrvl+2)
iQCur=limit(m_iMinQScale, (int)m_dQCur, m_iMaxQScale)
[0054] 3.6 QP Calculation for P-VOP
[0055] To calculate a target Qscale, the following five steps are
performed:
[0056] Step 1: calculate the average bit count (AvgTargetBitP)
obtained in the "bit allocation" module (using future frames) as
follows:
iAvgTargetBitP=(iTotalBudget/(dRCWindowSize+3.7)*RC_SPAN)
[0057] Step 2: calculate the previous bit count (ActualBitPrv)
(using past encoded frame)
[0058] Step 3: calculate the target Qscale using AvgTargetBit and
ActualBitPrv to adjust the deficit/surplus bit rate budget. If the
previous P-VOP spent 15% more than an average bit count, then QP
should be increased by its deviation. If the previous P-VOP spent
around 20% less than an average bit count, then QP should be
decreased by its deviation.
[0059] If none of these two conditions, then QP remains unchanged.
The description of this step is pseudo-encoded as follows:
3 if (iActualBitPrvP > iAvgTargetBitP*1.15) { dQCur = (iQPrv +
(iQPrv*(iActualBitPrvP- AvgTargetBitP)/iAvgTarge- tBitP)*0.55); }
else if (iActualBitPrvP < m_iAvgTargetBitP/1.20) { dQCur =
(iQPrv - (iQPrv*(iAvgTargetBitP-
iActualBitPrvP)/iAvgTargetBitP)*0.55); } else dQCur = iQPrv;
[0060] Step 4: refine Qscale using SAD statistics from past encoded
frames and current buffer status. Refining Qscale is conjunction
with buffer control and SAD statistics. In the case of
over-spending bit budget (i.e., case 1), the encoder has to
consider the potential buffer underflow problem by increasing
Qscale. The Qscale is scaled up by buffer deviation. Besides, to
further refine Qscale, SAD is used to determine the final Qscale.
If the deviation of the current SAD is larger than 10%, meaning a
more complex scene is on its way, thus the final Qscale should be
increased by the amount of iQPrv*(dBufDeviationP/dScale, where
iQPrv is the previous Qscale to take a early action before a buffer
underflow occurs. If the deviation of the current SAD is smaller
than about 10%, a less complex scene is expected, then the final
Scale can be adjusted by its buffer status and bounded by its
previous Qscale to maintain a certain degree of video quality. The
following pseudo-code describes the procedure to refine Qscale:
4 dSadDeviationP = (dSadCurP - dAvgSadCurP)/dAvgSadCurP;
dDifferenceP = (double)(m_iOccupancy - m_iInitialOccupancy); //
Case 1: potential buffer underflow!! if (dDifferenceP < 0.0) {
dScale = 1.5; dBufDeviationP = -1*dDifferenceP/iInitialOccupancy;
if (dSadDeviationP > 0.1 ) { dQCur = limit(iQPrv, (dQCur),
iQPrv+iQPrv*(dBufDeviationP/dScal- e)); } else if (dSadDeviationP
< -0.1) { dQCur = limit(iQPrv-iQPrv*(dBufDeviationP/dScale),
(m_dQCur), iQPrv); } else dQCur =
limit(iQPrv-iQPrv*(dBufDeviationP/dS- cale), (m_dQCur),
iQPrv+iQPrv*(dBufDeviationP/dScale)); }
[0061] In the case of under-spending bit budget (i.e., case 2), the
encoder has more room to maintain the video quality. In this case,
the SAD is used to scale down the Qscale by its deviation to the
average SAD values. If the deviation of the current SAD is larger
than about 10%, meaning a more complex scene is on its way, thus
the final Qscale will be bounded by the PrevQscale or less to
maintain constant quality since the encoder has more bits available
to spend. The final value of Qscale is decreased by the amount of
iQPrv*(dBufDeviationP/dScale, where iQPrv is the previous Qscale.
If the deviation of the current SAD is smaller than about 10%, a
less complex scene is expected, then the final Qscale can be
maintained to avoid sudden quality change, even though the change
is directed to quality improvement. If none of the above two
conditions holds, meaning a smooth scene is expected, the final
Qscale is adjusted based on its buffer states.
[0062] The following pseudo-code describes the procedure to refine
Qscale:
5 // case 2: easy mode for rate control else { dScale = 2.0;
dBufDeviationP = dDifferenceP/(double)m_iInitialO- ccupancy; if
(dSadDeviationP > 0.1) { dQCur =
limit(iQPrv-iQPrv*(dBufDeviationP/dScale), (m_dQCur), iQPrv); }
else if (dSadDeviationP < -0.1) { dQCur = limit(iQPrv,
(m_dQCur), iQPrv+iQPrv*(dBufDeviationP/dScale)); } else { dQCur =
limit(iQPrv-iQPrv*(dBufDeviationP/dS- cale), (m_dQCur),
iQPrv+iQPrv*(dBufDeviationP/dScale)); } }
[0063] Step 5: bound Qscale in the range of pre-defined MinQP and
MaxQP.
[0064] 4. Execution Phase 2: Encoding
[0065] In the encoding stage, if either the frame- or object-level
rate control can be activated, the encoder just simply encodes the
video frame or object using the value of QP obtained in the
pre-encoding stage. However, some low-delay applications may
require more strict buffer regulations, for example 250 ms for the
maximal accumulated delay, or higher bit rate encoding (e.g.,
1.about.4 Mbps encoding at CCIR-601 resolution), or
perceptual-based encoding, then a macroblock-level rate control is
expected. However, the macroblock level rate control is costly at
low rate since there is additional overhead if the quantization
parameter is changed within a frame. For example, in the MPEG-4
video, the MB (MacroBlock) type has to be encoded with three more
bits indicating the existence of the differential quantization
parameter (i.e., dquant).
[0066] Furthermore, two bits need to be sent for dquant as
described in MPEG-4 documentation. For the same prediction mode, an
additional 5 bits need to be transmitted in order to change QP. In
the case of encoding at 10 kbps, 7.5 fps, qcif resolution, the
overhead is computed as high as 99*5*7.5=3.7 kbps. If only 33
macroblocks are encoded, the overhead 33*5*7.5=1.2 kbps. Thus,
there will be about 10 percent loss in compression efficiency at
low bit rate encoding. At high bit rate, the overhead bit count is
less significant than the residual bit count.
[0067] 5. Execution Phase 3: Post-Encoding
[0068] In the post-encoding stages, the encoder simply records the
statistical data such as the bit usage, MAD or SAD and the
distortion of the decoded picture from its original picture. The
encoder also updates a VBV buffer status. Unlike the first order or
the second order rate distortion models that requires computational
load for model parameter derivation.
[0069] 5.2 VBV Buffer Update
[0070] To update the VBV buffer status, the channel input rate to
the VBV buffer is calculated based on the lapsed time. Then if the
VBV buffer fullness (i.e., iOccupancy) is smaller than the actual
coded size of the current frame, then a buffer underflow occurs. To
deal with buffer underflow, one solution in the present invention
is to freeze the previous frame by replacing the already encoded
bits with stuffing bits which basically assume all macroblocks are
marked as Skip macroblocks. In the case of buffer overflow, an easy
solution which involves stuffing more bits into the bitstream may
be utilized. The number of stuffing bits is given by:
iOccupancy-iOverFlowLevel
[0071] Wherein iOverFlowLevel is around the size of the VBV buffer.
The pseudo-code of this part is shown below:
6 if (iOccupancy <= (iActualBitCur+8) { // buffer underflow
replacing the coded bits of the current frame with stuffing bits }
else{ iNumberOfStuffingBits = iOccupancy - iOverFlowLevel; if
(iNumberOfStuffingBits > 0) do bits stuffing
(iNumberOfStuffingBits); else iNumberOfStuffingBits = 0; }
Occupancy += iChannelInputRate - iNumberOfStuffingBits
[0072] B. Scalable Macro-Block Rate Control
[0073] 1. Overview
[0074] The following discussion illustrates an example embodiment
of the present invention. It will be appreciated that labels,
terms, sequences and other specific parameters are used in
connection with the example described herein and may be changed
without departing from the scope of the invention.
[0075] 1.1 Introduction
[0076] The macro-block rate control (MbRc) methodology of the
present invention is intended for bit rate and quality control for
MPEG-4 encoders, such as the Sony MPEG4 AVC/ITU H.264 encoder. To
activate this MbRc, its picture level rate control must be enabled
to pass the target QP of a frame to MbRc. Note that the methodology
of the MbRc can be applied to any rate control scheme as long as
the Qp value for a picture is provided. There are two versions of
this implementation of MbRc. One is the floating-point
implementation for high performance computing system (e.g., Intel
IA-32 platform); the other is fixed point implementation for ARM
based platform without powerful floating-point capability. This new
MbRc is based on the notion of "intensity", it is very simple to
implement, and provides very effective performance in visual
quality. A flow diagram of the methodology is illustrated in FIG. 6
wherein the MB rate control steps are shown in blocks 100, 102,
106, 108, 112 and 114.
[0077] 1.2 Enabling Floating Point MB-RC
[0078] Floating point MB-RC is enabled as follows:
[0079] In StatisticalDefine.h file,
[0080] #define NEW_PIC_RC
[0081] #define _SRC_MB_ // turn on MB rate control
[0082] 1.3 Enabling Fixed Point MB-RC
[0083] Fixed point MB-RC is enabled as follows:
[0084] In StatisticalDefine.h file,
[0085] #define NEW_PIC_RC
[0086] #define _SRC_MB_ // turn on MB rate control
[0087] In JvtScalableRateControlMB.h
[0088] #define SRCMB_INT.sub.--
[0089] 1.4 Enabling Statistical Printout
[0090] Statistical printout is enabled as follows:
[0091] In JvtScalableRateControlMB.h
[0092] Prerequisite: _SRC_MB_ is enabled in StatisticDefine.h
[0093] if (defined(_SRC_MB_INFO_ONLY_) &&
[0094] defined(_SRC_MB_PRINTOUT_) then
[0095] MB info of PIC-level RC is recorded
[0096] if (undefined(_SRC_MB_INFO_ONLY_) &&
[0097] defined(_SRC_MB_PRINTOUT_) then
[0098] MB-level RC is enabled and MB info is recorded
[0099] //#define _SRC_MB_INFO_ONLY.sub.--
[0100] #define _SRC_MB_PRINTOUT.sub.--
[0101] 2. MB-RC Class Structure
[0102] 2.1 Data Members
[0103] The following are examples of data members used in this
embodiment of this invention:
7 JvtRcParameter* m_pRcParameter; // Pointer of JvtRcParameter
class JvtPicParam* m_pPicParam; // Pointer of JvtPicParam class
JvtMbParam* m_pMbParam; // Pointer of JvtPicParam class int
m_iMB_h; // number of MBs in x-axis int m_iMB_v; // number of MBs
in y-axis int m_iNumOfMb; // total number of MBs int
m_iNumofCodedMb; // total number of coded MBs int m_iPicTypeCur; //
picture type of the current frame int m_iRetryCount; // retry
counter int m_iMbAccmQp; // accumulated syntax Qps int m_iMbQpCur;
// syntax Qp of current MB int m_iMbQpPrv; // syntax Qp of previous
MB int m_iPicEstSyntaxQp; // estimated Qp from PicRc int
*m_piMbMad; // a pointer to a Mad map int m_iMbAvgMad; // average
Mad of a slice(frame) int m_iMbAvgAd; // average Ad of a
slice(frame) int m_iMbTotalAd; // total Ad of a slice int
*m_piMbDistortionMap; // a pointer to a distortion map int
m_iMbIntraAdCur; // IntraAd of current MB int m_iMbInterAdCur; //
InterAd of current MB int m_iMbIntraQp; // IntraQp of current MB
int m_iMbInterQp; // InterQp of current MB #ifdef
_SRC_MB_INT.sub.-- int m_iQpScale; int m_iQpConverge; int
m_iAdjustStepSize; int m_iMbAvgDistortion; int m_iPicEstActualQp;
static const int m_iQpActualToSyntaxTable[365]; #else //
_SRC_MB_INT.sub.-- double m_dQpScale; // Qp scalor to converge MbQp
to PicQp double m_dQpConverge; // converge speed double
m_dAdjustStepSize; // QpStep granularity double m_dMbAvgDistortion;
// Average distortion double m_dPicEstActualQp; // Estimated Actual
Pic Qp #endif//_SRC_MB_INT.sub.--
[0104] 2.2 Member Functions
[0105] The following are examples of the member functions employed
in the present invention:
[0106] int iecInit(JvtRcParameter *pRcParameter, JvtEtcParameter
*pEtcParameter);
[0107] int iRateCtrlMbBootUp(JvtRcParameter *pRcParameter,
JvtPicParam* pJvtPicParam, int iOverheadBits);
[0108] int iRateCtrlMbCalcComplexitylonly(int iMbNum, JvtPicParam*
pPicParam, int iIntraAd);
[0109] int iRateCtrlMbCalcQp(int iMbNum, JvtMbParam* pMbParam, int
iIntraAd, int iInterAd);
[0110] int iRateCtrlMbUpdate(int iMbNum, JvtMbParam* pMbParam, int
iPicActBit);
[0111] Int iRateCtrlMbAfter( );
[0112] 3. MB-RC Implementation
[0113] Various parameters and functions associated with the
methodology of the present invention are described as follows:
[0114] 3.1 MB-RC iecInit
[0115] The member function iecInit(JvtRcParameter *pRcParameter,
JvtEtcParameter *pEtcParameter) creates a working memory, and
initializes slice-wide variables including: m_iMB_h, m_iMB_v,
m_iNumOfMb and m_iQpConverge.
[0116] 3.2 MB-RC Boot Up
[0117] The member function iRateCtrlMbBootUp(int iRetryCount,
JvtPicParam* pJvtPicParam, int iOverheadBits) is called before
encoding a slice in JvtSliceEncoding before the MB-coding loop.
This function generally performs content complexity analysis and
initializes slice-wide variables. Note that iRetryCount indicates
the number of occurrences of re-encoding the current slice due to
the VBV underflow. Normally it comprises a zero value. The MB-RC
Boot Up process proceeds according to the following steps.
[0118] Step 1: Calculate the MAD of each MB, comprising 4 blocks,
as a basis for content complexity estimation, from the following: 1
MbMad = ( MAD of the original source frame ) per pixel ( int ) ( (
i = 0 i = 3 j = 0 j = 63 X i , j - X _ i ) / 256.0 )
[0119] where X.sub.i,j denotes the pixel value at position j at
block i, and {overscore (X)}.sub.i denote the mean value of the
block i.
[0120] Step 2: Calculate the average MAD of the entire frame or
slice from the following: 2 MbAvgMad = ( { n = 0 n = k [ ( int ) (
( i = 0 i = 3 j = 0 j = 63 X i , j - X _ i ) / 256.0 ] } / k )
[0121] where k is the total number of MBs.
[0122] Step 3: Determine the intensity base of the entire
frame/slice as follows:
IntensityBase=(MbAvgMad+1)/2
[0123] where IntensityBase is used as a bias to avoid large
variation of intensity due to small MAD values.
[0124] Step 4: Determine the Qp dynamic range (i.e., LocalAdjRange)
of the entire frame/slice as follows:
LocalAdjRange=min((51-SyntaxQp)*(SyntaxQp+200)/6400.0,
MAX_RANGE),
[0125] wherein MAX_RANGE=0.3 in this implementation. Certainly this
number is controllable, depending on the scene content of a
sequence. This value can be changed either in sequence base or in
frame base. The larger its value is, the broader the Qp dynamic
range is. If MAX_RANGE=0.0, then the Qp of the underlying picture
level is used.
[0126] 3.3 MB-RC Qp Calculation
[0127] 3.3.1 I-Slice Coding
[0128] For I-slice:
[0129] iRateCtrlMbCalcComplexitylonly(int iMbNum, JvtPicParam*
pPicParam, int iIntraAd)
[0130] Step 1: Obtain iMbIntraAdCur from Mode decision module and
divide IntraAd by 256 for per pixel basis.
[0131] Step 2: Calculate various intensities. Here, two separate
cases are considered: Case 1 is to encode the first I-frame where
there is no average Ad and distortion information is available
except the retrying encoding (then average Ad and distortion
information is available, and Case 2 can be applied). Case 2 is a
regular I-frame, which is inserted in every certain pre-specified
interval.
[0132] Case 1: Scene Change I Slice or the First I Slice given no
retry encoding:
MadIntensity=(MbMad[iMbNum]+1)/(MbAvgMad+1);
AdIntensity=(MbInterAdCur+1)/(EstMbAvgAd+1);
[0133] Where EstMbAvgAd=(m_iMbAvgMad+1)/MAD_DIV_AD_PER_PIXEL, and
MAD_DIV_AD_PER_PIXEL=1.5 in this implementation based on the
empirical values in Tables 4 and 5 which show the MAD and AD value
in the selected sequences, and their relation.
[0134] Next, the index is calculated, T1, representing coding
(i.e., Ad) activity, and limit T1 in +/- LocalAdjRange that is
calculated in step 4 of section 3.2 to avoid large
fluctuations.
T1=MadIntensity/AdIntensity
[0135] These indexes provide a measure of the intensity of Mad
versus Ad. The basic empirical observation is to assume that if T1
equals to 1, meaning that Ad and Mad both are in average level, the
Qp level should be kept at the Pic target Qp. If T1 is greater than
1, meaning the content of this MB is pretty busy (above average),
its coding complexity (AD) is below average, so its QP will be
increased due to its high complex content. On the other hand, if T1
is smaller than 1, then Qp will be decreased due to its "easy", or
flat content.
8 If (MadIntensity <= 1.0) { T1 = MIN(T1, MadIntensity) } else {
T1 = MAX(T1, MadIntensity) }
[0136] To weigh the content complexity of a MB, we also add the
above MIN and MAX operations to ensure the MB is properly
interpreted.
[0137] Then limiting IntraLocalAdj:
[0138] // coding intensity
IntraLocalAdj=limit(1-LocalAdjRange, T1, 1+LocalAdjRange);
[0139] Case 2: Regular I-Frame(Slice). In addition to the above
similar calculations for Case 1, an additional index T2 is needed.
The new intensity calculations are listed below:
DistIntensity=(MbDistortionMap[iMbNum]+IntensityBase)/(MbAvgDistortion+Int-
ensityBase)
MadIntensity=(MbMad[iMbNum]+IntensityBase)/(MbAvgMad+IntensityBase)
AdIntensity=(MbInterAdCur+IntensityBase)/(MbAvgAd+IntensityBase);
[0140] where IntensityBase is derived in step 3 of section 3.2.
[0141] Next, indices, T1, and T2, are calculated which represent
coding (i.e., Ad) activity and Distortion versus Mad, and limit T1
and T2 in +/- LocalAdjRange calculated in step 4 of section 3.2 to
avoid large fluctuations.
9 T1 = MadIntensity / AdIntensity T2 = MadIntensity / DistIntensity
Then, If (MadIntensity <= 1.0) { T1 = MIN(T1, MadIntensity) T2 =
MIN(T2, MadIntensity) } else { T1 = MAX(T1, MadIntensity) T2 =
MAX(T2, MadIntensity) }
[0142] Then limiting T1 and T2:
[0143] // coding intensity
T1=limit(1-LocalAdjRange, T1, 1+LocalAdjRange);
[0144] // distortion intensity
T2=limit(1-LocalAdjRange, T2, 1+LocalAdjRange);
[0145] Next, combining both indexes into one by weighting each
index as:
IntraLocalAdj=X*T1+(1.0-X)*T2
[0146] wherein X is an empirical value, and can be controlled by
the application, such as given by X=0.5 in this implementation.
[0147] From this point, both cases execute the following steps.
[0148] Step 3: determine the QpStep granularity as:
IntraLocalAdjust=(IntraLocalAdjust-1.0)*AdjustStepSize.
[0149] Note that AdjustStepSize is decided by the content
complexity of the current MB. AdjustStepSize is a key factor to
determine the range of Qp dynamic changes and is derived as: 3
dAdjustStepSize = ( iMbQpGap - m_iPicEstSyntaxQp ) * 0.01 and
iMbQpGap = 200 if m_piMbMad [ iMbNum ] < m_iMbAvgMad ; = 80 if (
m_piMbMad [ iMbNum ] > m_iMbAvgMad ) = 130 otherwise .
[0150] Note that these numbers (200, 80 and 130) are empirical
numbers, and can be modified as needed. The value m_iMbAvgMad is
the average Mad per pixel and m_piMbMad[iMbNum] is the Mad value
per pixel of the current MB, iMbNum. The detail is described in
section 3.2. It should be appreciated that as the value of
AdjustStepSize increases, the Qp dynamic range increases.
[0151] Step 4: IntraLocalAdjust is offset by the average local
adjust of its previous slice(frame) in order to be closer to the
target picture Qp.
IntraLocalAdj-=AvgLocalAdj
[0152] Step 5: calculate the first actual
Qp=m_dPicEstActualQp*(1+IntraLoc- alAdj)
[0153] Step 6: convert to the syntax m_iMbQpCur from
(int)(6*log(Qp)/log(2)+0.5).
[0154] Step 7: smooth iMbQpCur using a simple linear filter
m_iMbQpCur=(int)(m_iMbQpPrv*SRC_MB_PRVMB_EFFECT+m_iMbQpCur*(1.0-SRC_MB_PRV-
MB_EFFECT))
[0155] where iMbQpPrv is the syntax Qp of the previous MB, and
SRC_MB_PRVMB_EFFECT is the weighting factor of the previous MB. In
this implementation, its value is 0.2.
[0156] Step 8: finally the iMbQpCur is capped between
SRC_MB_MIN_SYNTAX_QP and SRC_MB_MAX_SYNTAX_QP.
[0157] 3.3.2 P-Slice Coding
[0158] For P-slice and B-slice:
[0159] iRateCtrlMbCalcQp(intiMbNum, JvtMbParam* pMbParam, int I
IntraAd, int iInterAd)
[0160] The following discussion is for a inter macro-block only.
For a inter macro-block, all of the execution steps are very
similar, except changing Inter MB to Intra MB.
[0161] Step 1: Obtain IntraAd and InterAd from Mode decision module
and divide IntraAd and InterAd by 256 for per pixel basis.
[0162] Step 2: Calculate Distortion, Mad and Ad intensity as
follows:
DistIntensity=(MbDistortionMap[iMbNum]+IntensityBase)/(MbAvgDistortion+Int-
ensityBase)
MadIntensity=(MbMad[iMbNum]+IntensityBase)/(MbAvgMad+IntensityBase)
AdIntensity=(MbInterAdCur+IntensityBase)/(MbAvgAd+IntensityBase)
[0163] Note that the DistortionMap and its average is derived from
its previous frame (slice) since we assume that the content
activity between two successive frames (except scene cut) is very
similar in co-location MB. And the AvgAd is obtained from its
previous frame (slice) too.
[0164] Step 3: Calculate two indexes, T1 and T2, representing
coding (i.e., Ad) and content (i.e., Distortion) activities, and
limit T1 and T2 in +/- LocalAdjRange in section 3.2 to avoid large
fluctuations, where its maximum value is 0.3 in this example. If
LocalAdjRange=0.0, then the picture level Qp is used.
T1=MadIntensity/AdIntensity
T2=MadIntensity/DistIntensity
[0165] These indexes provide a measure of the intensity of Mad
versus Ad, and distortion. The basic empirical observation is to
assume that if T1 equals to 1, meaning that Ad and Mad both are in
average level, the Qp level should be kept at the Pic target Qp. If
T1 is large than 1, meaning the content of this MB is pretty busy
(above average), its coding complexity (AD) is below average, so
its QP will be increased due to its high complex content. On the
other hand, if T1 is smaller than 1, then Qp will be decreased due
to its "easy", or flat content.
10 If (MadIntensity <= 1.0) { T1 = MIN(T1, MadIntensity) T2 =
MIN(T2, MadIntensity) } else { T1 = MAX(T1, MadIntensity) T2 =
MAX(T2, MadIntensity) }
[0166] To weigh the content complexity of a MB, we also add the
above MIN and MAX operations to ensure the MB is properly
interpreted.
[0167] Step 4: combine both indexes into one by weighting each
index as:
InterLocalAdjust=X*T1+(1.0-X)*T2
[0168] wherein X is an empirical value, and can be controlled by
the application, such as given by X=0.5 in this implementation.
[0169] Step 5: determine the QpStep granularity as
(InterLocalAdjust-1.0)*AdjustStepSize.
[0170] Next we decide the Qp directions (increasing/decreasing)
depending on the result of step 5. If it is less than 1, then Qp
will be decreased by (InterLocalAdjust-1.0) times AdjustStepSize.
AdjustStepSize is a key factor to determine the range of Qp dynamic
changes and is derived as: 4 dAdjustStepSize = ( iMbQpGap -
m_iPicEstSyntaxQp ) * 0.01 and iMbQpGap = 200 if m_piMbMad [ iMbNum
] < m_iMbAvgMad ; = 80 if ( m_piMbMad [ iMbNum ] >
m_iMbAvgMad ) = 130 otherwise .
[0171] Note that these numbers (200, 80 and 130) are empirical
values, and can be modified as needed. The value m_iMbAvgMad is the
average Mad per pixel and m_piMbMad[iMbNum] is the Mad value per
pixel of the current MB, iMbNum. The detail is described in section
3.2. The larger the value of AdjustStepSize, the greater the Qp
dynamic range.
[0172] Step 6: InterLocalAdjust is offset by the average local
adjust of its previous slice(frame) in order to be closer to the
target picture Qp.
IntraLocalAdj-=AvgLocalAdj
[0173] Step 7: calculate the first actual
Qp=m_dPicEstActualQp*(1+dIntraLo- calAdj)
[0174] Step 8: convert to the syntax m_iMbQpCur from
(int)(6*log(Qp)/log(2)+0.5)
[0175] Step 9: smooth iMbQpCur using a simple linear filter
m_iMbQpCur=(int)(miMbQpPrv*SRC_MB_PRVMB_EFFECT+m_iMbQpCur*(1.0-SRC_MB_PRVM-
B_EFFECT)),
[0176] where iMbQpPrv is the syntax Qp of the previous MB, and
SRC_MB_PRVMB_EFFECT is the weighting factor of the previous MB. In
this example, its value is 0.2.
[0177] Step 10: finally the iMbQpCur is capped between
SRC_MB_MIN_SYNTAX_QP and SRC_MB_MAX_SYNTAX_QP.
[0178] 3.4 MB-RC Update
[0179] iRateCtrlMbUpdate(int iMbNum, JvtMbParam* pMbParam, int
iPicActBit)
[0180] This function provides for collecting the actual AD
(depending on MB-type), Qp, and determining the QpScale value for
the next MB coding.
[0181] 3.5 MB-RC Clean UP
[0182] iRateCtrlMbAfter( )
[0183] This function is to calculate the distortion of a MB, and
sum up for a frame. It also calculates the average Ad of a frame.
The MB distortion is its Sum of Absolute Difference (SAD).
[0184] 4. Details of MB Rate Control Steps
[0185] 4.1 MB Rate Control Execution Flow and Data Flow
[0186] Referring again to FIG. 6, the execution flow and data flow
according to an embodiment of the invention is illustrated as
follows.
[0187] Block 100 illustrates the iecInit step which comprises the
following:
[0188] (a) initialize data members;
[0189] (b) set image dimension information is set; and
[0190] (c) allocate working memory.
[0191] Block 102 illustrates the iRateCtrlMbBootUp step which
comprises the following:
[0192] (a) Obtain the picture level target Qp from picture level
rate control;
[0193] (b) Calculate each MB content complexity (MbMad) and average
Mad;
[0194] (c) Determine intensity base; and
[0195] (d) Determine local adjust range.
[0196] Block 104 illustrates the Mode Decision step which comprises
calculating InterAD and IntraAD. If the steps at blocks 106 and 108
have already been executed, the mode decision module determines the
final mode and its Qp.
[0197] Block 106 illustrates the iRateCtrlMbCalcComplexitylonly
step which comprises the following:
[0198] (a) Obtain IntraAd from mode decision module;
[0199] (b) Calculate MB-level Qp in I-slice; and
[0200] (c) Send this Qp back to mode decision module.
[0201] Block 108 illustrates the iRateCtrlMbCalcQp which comprises
the following:
[0202] (a) Obtain IntraAd and InterAd from mode decision
module;
[0203] (b) Calculate two MB-level Qps, IntraQp and InterQp, for
Inter mode and Intra mode, respectively in P-slice; and
[0204] (c) Send back these two Qps to mode decision module.
[0205] Block 110 illustrates the iecMbCoding step which comprises
encoding the current MB.
[0206] Block 112 illustrates the vRateCtrlMbUpdate step which
comprises the following:
[0207] (a) Accumulate Qp, Ad and the number of coded MBs; and
[0208] (b) If no mode MB coding, jump to block 104, else go to
block 114.
[0209] Block 114 illustrates the vRateCtrlMbAfter step which
comprises the following:
[0210] (a) Calculate the distortion of the current frame;
[0211] (b) Calculate the average Ad for the next frame; and
[0212] (c) Calculate the average LocalAdj for the next frame.
[0213] 4.2 Content Complexity Representation
[0214] Content complexity analysis, per MB Mad calculation and its
average value, is performed before the first MB of a slice is
encoded. The calculation of per MB Mad and its average is shown
below.
[0215] (a) Calculating the MAD of Each MB, Comprising 4 Blocks, as
a Basis for Content Complexity Estimation. 5 MbMad = ( MAD of the
original source frame ) per pixel ( int ) ( ( i = 0 i = 3 j = 0 j =
63 X i , j - X _ i ) / 256.0 )
[0216] Wherein X.sub.i,j denotes the pixel value at position j at
block i, and {overscore (X)}.sub.i denote the mean value of the
block i.
[0217] (b) Calculating the Average MAD of the Entire Frame or
Slice: 6 MbAvgMad = ( { n = 0 n = k [ ( int ) ( ( i = 0 i = 3 j = 0
j = 63 X i , j - X _ i ) / 256.0 ] } / k )
[0218] Wherein k is the total number of MBs.
[0219] 4.3 Intensity Based QP Adjustment
[0220] The present invention further comprises a new intensity
based approach to derive the Qp value for each MB. The intensity is
defined as (current_value)/(average_value). In this implementation,
three different intensities are used; namely Mean Absolute
Difference (MAD) intensity, Absolute Difference (AD) intensity, and
Distortion intensity. Furthermore, to reduce the "noise" effect on
this calculation, an intensity base is added, which is used to
reduce this effect to avoid large Qp fluctuation. In this
implementation, the intensity based is defined as (average
MAD)/2.
[0221] The three intensities according to the present invention are
defined as:
DistIntensity=(MbDistortionMap[iMbNum]+IntensityBase)/(MbAvgDistortion+Int-
ensityBase)
MadIntensity=(MbMad[iMbNum]+IntensityBase)/(MbAvgMad+IntensityBase)
AdIntensity=(MbInterAdCur+IntensityBase)/(MbAvgAd+IntensityBase);
[0222] Note that the MAD and AD calculations were previously
defined at paragraph [0084]. Basically MAD intensity denotes
content complexity, and AD denotes the coding complexity since it
is obtained from motion estimation module. Distortion intensity
provides compensation for the coding artifact.
[0223] Then two indices, T1 and T2, are introduced to measure the
relative intensity, as shown below
T1=MadIntensity/AdIntensity
T2=MadIntensity/DistIntensity
[0224] Generally speaking, smaller T1 and T2 shall derive smaller
Qp, resulting in better quality. The basic empirical observation is
to assume that if T1 equals to 1, meaning that Ad and Mad both are
in average level, Qp level should be kept at the Pic target Qp. If
T1 is greater than 1, meaning the content of this MB is pretty busy
(above average), but its coding complexity (AD) is below average,
its QP will be increased due to its high complexity content. On the
other hand, if T1 is smaller than 1, then Qp will be decreased due
to its "easy", or flat content.
[0225] To further differentiate the content of a MB, a simple
maximum and minimum operation can be used to select the larger
value or smaller value, respectively.
11 If (MadIntensity <= 1.0) { T1 = MIN(T1, MadIntensity) T2 =
MIN(T2, MadIntensity) } else { T1 = MAX(T1, MadIntensity) T2 =
MAX(T2, MadIntensity) }
[0226] For example, if two macro-blocks (M1, and M2) both have the
same T1(M1)=T1(M2)=1.0, but MadIntensity(M1)=2.0 and
MadIntensity(M2)=1.0. Then without this MIN and MAX operation, both
will have the same T1 value. However, with this operation,
T1(M1)=1.0, but T1(M2) becomes 2.0, resulting in higher Qp. This
result is expected since any noise in higher content complexity is
less sensitive to this vision system.
[0227] Finally, T1 and T2 are limited in +/-LocalAdjRange range,
where LocalAdjRangew will be explained in section [00161].
T1=limit(1-LocalAdjRange, T1, 1+LocalAdjRange);
T2=limit(1-LocalAdjRange, T2, 1+LocalAdjRange)
[0228] Note that in the first I-frame(slice) and scene change
I-frame(slice), the distortion information is not available before
encoding a MB, then only T1 will be derived. Otherwise both T1 and
T2 have to be calculated.
[0229] 4.4 Adaptively Adjust Qp Range
[0230] LocalAdjRange is calculated as follows:
MIN((51-SyntaxQp)*(SyntaxQp+200)/6400.0, MAX_RANGE)
[0231] where MAX_RANGE=0.3 in this implementation. This
LocalAdjRange is controllable, depending on the scene content of a
sequence. This value can be changed either in sequence base or in
frame base. The larger its value, the broader the Qp dynamic range.
If MAX_RANGE=0.0, then the Qp of the underlying picture level is
used. LocalAdjRange also shows that in the higher Qp in picture
level, the dynamic range of Qp is narrowed, and its distribution
versus SyntaxQp is depicted in FIG. 7.
[0232] 4.5 Adaptively Adjust Qp Step Size
[0233] After determining T1 and T2 in section [00151], the next
step is to determine the Qp adjustment, LocalAdj. The formula is as
follows:
LocalAdj=limit(1-LocalAdjRange, AD_WEIGHT*T1+(1-AD_WEIGHT)*T2,
1+LocalAdjRange),
[0234] where LocalAdjRange is described in section [00161], and
AD_WEIGHT is the fraction value in the range of 0.0 to 1.0. In our
case, 0.5 is used to show that both T1 and T2 are equally
important. LocalAdj is in the range of 1+/-LocalAdjRange. Smaller
LocalAdj (i.e., <1.0) will have smaller Qp (i.e.,
<picture-level Qp).
[0235] Before going into the detailed discussion of "Adjust Qp Step
size", we review the process of Qp calculation.
[0236] Step 1: LocalAdj=(LocalAdj-1.0)*AdjustStepSize;
[0237] Step 2: LocalAdj=AvgLocalAdj;
[0238] Step 3: Qp=PicEstActualQp*(1+LocalAdj);
[0239] Step 1 is to transform the LocalAdj from the range of
(1+/-LocalAdjRange) to (+/-LocalAdjRange)* AdjustStepSize. That is,
LocalAdj will increase (i.e., positive) or decrease (i.e.,
negative) the Qp value because of Step 3. Step 2 is performed to
compensate the discrepancy between picture level target Qp, and MB
level actual Qp in the previous frame (slice). Ideally the sum of
LocalAdj for all MBs in a frame (slice) should be 0.0, so any
leftover of sum of LocalAdj will be propagated to the next frame
(slice) to ensure overall its approximation of picture-level Qp and
bit rate.
[0240] Now let's discuss AdjustStepSize in step 1. This variable is
to determine the granularity of adjustment step size. Its
implementation is described below:
12 if (MbMad[iMbNum] < MbAvgMad) { MbQpGap = 200; } else if
(MbMad[iMbNum] > MbAvgMad) { MbQpGap = 80; } else { MbQpGap =
130; } AdjustStepSize = (iMbQpGap-PicEstSyntaxQp)*0.01.
[0241] Note that this is just an example of calculation of
AdjustStepSize. These numbers 200, 80 and 130 are empirical values,
and can be modified according to different encoding environments.
The basic idea of this method is to control the AdjustStepSize by
changing the MbQpGap based on its content complexity. This
implementation says that if the content of current MB, (i.e.,
MbMad[iMbNum]), is less than average complexity (MbAvgMad), then we
can enlarge the MbQpGap value to significantly improve the
perceptual quality in this smooth area. On the other hand, if it is
larger than average complexity, a narrower MbQpGap is used to
preserve the quality of this complex area. Flat areas show bigger
MbQpGap values, resulting in significant improvement in visual
quality, while in complex or busy areas, MbQpgap becomes small to
preserve the quality to some degree. Those MBs with a negative
value will be assigned smaller Qp and those MBs with positive
values will be assigned bigger Qp for step 3.
[0242] 4.6 Approximation of Picture Level Qp
[0243] In the MB algorithm of the present invention, there is no
target bit information from picture level rate control. Instead,
only the picture level target Qp is given to MB rate control.
Therefore, to properly control the bit rate (to make both of them
generate around the same bits in picture level and sequence level),
the MB rate control is trying to get the average Qp, which is about
the same value as PIC rate control. The way the MB rate control
does this is to calculate the average Qp, and also calculate the
average LocalAdj to realize that if the current local adjustment is
too light or too overdosing, then the leftover part will be
propagated to the next frame. For example, if the average
LocalAdjust is -0.22, meaning that in this frame, most of the MBs
will have smaller Qp than picture level Qp, then in the next frame,
we have to compensate it by passing this information to the next
frame. Thus, when MB rate control calculates the local adjust for
the next frame, this leftover part will be taken into account in
step 2 of section [00163].
[0244] 5. Experimental Results
[0245] 5.1 Target Bit Rate Coding
[0246] Tables 6 and 7 illustrate the performance of the macroblock
rate control algorithm described above versus picture level rate
control. Table 6 shows the performance of MB rate control, while
Table 7 shows the performance of picture level rate control.
[0247] The first column in both tables shows the coding condition
of a bitstream. It is specified as follows, for example,
[0248] bicy.sub.--1000K.sub.--30F_G2_D31_Db1_ep0
[0249] where bicy: sequence name, 1000K: target bit rate is 100
Kbits per second, 30F: target frame rate is 30 frames per second,
G2: insert I-frame is every two seconds, D31: delay is 31 frames,
Db1: deblocking filter is enabled, ep0: CAVLC entropy coding is
used. The second column in both tables shows the dimension of the
image and its scan mode: i denotes interlaced source video, and p
denotes progressive source video. The third column (R) shows the
actual bit rate. The third column (I-Qp) shows average Qp for 1
frames. The fourth column (P-Qp) shows the average Qp for P frames.
The seventh to ninth columns show the average PSNR values.
[0250] In this experiment, MbRc was slightly worse in terms of PSNR
value, but in terms of visual quality, it demonstrated a
significantly superior visual quality due to its intensity based
bit distribution. The basic idea is to lower Qp value to
significantly improve the flat and smooth area, to which the human
visual system is quite sensitive, while it increases Qp value in
busy and complex areas (insensitive to the vision system) to
improve the overall bit rate and quality.
[0251] 5.2 Fix QP Coding
[0252] Table 8 shows all the frame coding information including Qp,
Bits and PSNR for picture level rate control and MB level rate
control in encoding a sample sequence using Qp=35. The last row
summarizes the results that picture level rate control spends 50107
bits on average per frame, while the MB rate control of the present
invention spends 42365 bits per frame and its average Qp=34.66. The
PSNR value is around 1.0 db difference. But in terms of the visual
quality, again the MB rate control shows around the same or better
quality than picture level rate control.
[0253] Although the description above contains many details, these
should not be construed as limiting the scope of the invention but
as merely providing illustrations of some of the presently
preferred embodiments of this invention. Therefore, it will be
appreciated that the scope of the present invention fully
encompasses other embodiments which may become obvious to those
skilled in the art, and that the scope of the present invention is
accordingly to be limited by nothing other than the appended
claims, in which reference to an element in the singular is not
intended to mean "one and only one" unless explicitly so stated,
but rather "one or more." All structural and functional equivalents
to the elements of the above-described preferred embodiment that
are known to those of ordinary skill in the art are expressly
incorporated herein by reference and are intended to be encompassed
by the present claims. Moreover, it is not necessary for a device
or method to address each and every problem sought to be solved by
the present invention, for it to be encompassed by the present
claims. Furthermore, no element, component, or method step in the
present disclosure is intended to be dedicated to the public
regardless of whether the element, component, or method step is
explicitly recited in the claims. No claim element herein is to be
construed under the provisions of 35 U.S.C. 112, sixth paragraph,
unless the element is expressly recited using the phrase "means
for."
13TABLE 1 iNumberOfMB = (iHSize / iYMBHSize) * (iVSize /
iYMBVSize); dBit rateForABlock = (m_pEnc->m_iBit rate) /
(iNumberOfMB) dFrameRate bit Bits per Bits per New rate framerate
width height MB pixel Ratio MinQP 10000 5 176 144 20.20 0.08 101.38
5 10000 7.5 176 144 13.47 0.05 152.06 6 10000 10 176 144 10.10 0.04
202.75 6 10000 15 176 144 6.73 0.03 304.13 8 10000 30 176 144 3.37
0.01 608.26 8 32000 10 176 144 32.32 0.13 63.36 4 32000 15 176 144
21.55 0.08 95.04 5 32000 30 176 144 10.77 0.04 190.08 6 64000 10
176 144 64.65 0.25 31.68 3 64000 15 176 144 43.10 0.17 47.52 3
64000 30 176 144 21.55 0.08 95.04 5 192000 10 352 288 48.48 0.19
42.24 3 192000 15 352 288 32.32 0.13 63.36 4 192000 30 352 288
16.16 0.06 126.72 6 256000 10 352 288 64.65 0.25 31.68 3 256000 15
352 288 43.10 0.17 47.52 3 256000 30 352 288 21.55 0.08 95.04 5
384000 10 352 288 96.97 0.38 21.12 1 384000 15 352 288 64.65 0.25
31.68 3 384000 30 352 288 32.32 0.13 63.36 4 1500000 30 1208 1152
9.20 0.04 222.66 8 2000000 30 1208 1152 12.26 0.05 166.99 6 4000000
30 1208 1152 24.53 0.10 83.50 5 6000000 30 1208 1152 36.79 0.14
55.66 4 8000000 30 1208 1152 49.06 0.19 41.75 3
[0254]
14TABLE 2 Positive part (I.e., buffer could be overflow) Prv Tar
Final Prv Tar Final QP QP Dev* 1-Dev QP QP QP Dev* 1-Dev QP 28 8 0
1 18 3 8 0 1 8 28 8 0.1 0.9 18 3 8 0.1 0.9 7.05 28 8 0.2 0.8 18 3 8
0.2 0.8 6.2 28 8 0.3 0.7 17.8 3 8 0.3 0.7 5.45 28 8 0.4 0.6 15.2 3
8 0.4 0.6 4.8 28 8 0.5 0.5 13 3 8 0.5 0.5 4.25 28 8 0.6 0.4 11.2 3
8 0.6 0.4 3.8 28 8 0.7 0.3 9.8 3 8 0.7 0.3 3.45 28 8 0.8 0.2 8.8 3
8 0.8 0.2 3.2 28 8 0.9 0.1 8.2 3 8 0.9 0.1 3.05 28 8 1 0 8 3 8 1 0
3
[0255]
15TABLE 3 Prv Tar Final Prv Cur Final QP QP Dev** QP QP QP Dev** QP
28 8 0 8 3 8 0 3 28 8 0.1 8.2 3 8 0.1 3.05 28 8 0.2 8.8 3 8 0.2 3.2
28 8 0.3 9.8 3 8 0.3 3.45 28 8 0.4 11.2 3 8 0.4 3.8 28 8 0.5 13 3 8
0.5 4.25 28 8 0.6 15.2 3 8 0.6 4.8 28 8 0.7 17.8 3 8 0.7 5.45 28 8
0.8 20.8 3 8 0.8 6.2 28 8 0.9 24.2 3 8 0.9 7.05 28 8 1 28 3 8 1
8
[0256]
16TABLE 4 SD Bit rate MAD per pixel AD per pixel MAD/AD bicycle
1000K 12.67 9.64 1.31 Bus 1000K 13.23 9.27 1.43 Car 1000K 11.28
8.21 1.37 cheer 1000K 12.53 8.68 1.44 confe 1000K 10.74 7.19 1.49
football 1000K 8.70 6.32 1.38 flower 1000K 14.84 11.32 1.31 marble
1000K 12.41 4.42 2.81 mobile 1000K 17.39 13.35 1.30 Pop 1000K 8.57
6.75 1.27 tennis 1000K 15.73 14.51 1.08 Face 1000K 2.82 0.79 3.57
1000K 3.24 1.75 1.85 1000K 3.54 2.22 1.59 Wall 1000K 1.69 1.02 1.66
1000K 3.03 1.59 1.91 1000K 2.03 1.38 1.47 battle 1000K 2.77 2.23
1.24 1000K 3.77 3.74 1.01 1000K 5.45 3.99 1.37 AVERAGE 8.32 5.92
1.59
[0257]
17TABLE 5 SIF/CIF Bit rate MAD per pixel AD per pixel MAD/AD
bicycle 384K 12.38 9.20 1.35 cheer 384K 15.07 11.82 1.27 child 384K
9.04 5.60 1.61 foreman 384K 6.68 3.50 1.91 fountain 384K 14.21
11.98 1.19 mino 384K 8.76 4.98 1.76 mobile 384K 18.76 13.94 1.35
new2 384K 7.70 4.84 1.59 papa 384K 3.79 2.57 1.47 AVERAGE 10.71
7.60 1.50
[0258]
18TABLE 6 PSNR- PSNY- PSNR- Bistream (MB) Dimension Bit rate I-Qp
P-Qp Y Cb Cr bicy_1000K_30F_G2_D31_Db1 _ep0 720 .times. 480i
1040.0K 42.28 43.00 23.24 33.75 32.76 bus_1000K_30F_G2_D31_Db1_ep0
720 .times. 480i 953.4K 38.65 39.26 25.66 37.40 35.60
cheer_1000K_30F_G2_D31_Db1_ep- 0 720 .times. 480i 1053.7K 43.77
45.17 21.90 30.17 29.27 child_384K_30F_G2_D31_Db1_ep0 352 .times.
288p 379.1K 32.43 33.67 31.62 33.93 33.63
flower_1000K_30F_G2_D31_Db1_ep0 720 .times. 480i 985.5K 39.44 40.54
22.91 32.18 30.92 foreman_384K_30F_G2_D31_Db1_e- p0 352 .times.
288p 385.9K 28.32 29.72 33.92 41.20 39.67
foundtain_1000K_30F_G2_D31_Db1_ep0 320 .times. 240p 1184.1K 32.76
35.54 26.47 44.12 43.35 mino_384K_30F_G2_D31_Db1_ep0 320 .times.
240p 342.2K 26.60 27.85 35.17 38.88 37.08
news2_384K_30F_G2_D31_Db1_ep0 320 .times. 240p 418.8K 28.30 30.29
34.26 39.24 37.74 papa_384K_30F_G2_D31_Db1_ep0 320 .times. 240p
380.8K 22.87 23.63 39.76 44.02 42.87 pop_1000K_30F_G2_D31_Db1_ep0
720 .times. 480i 1168.4K 37.40 38.86 27.55 32.52 32.07
sasam_384K_30F_G2_D31_Db1_ep0 320 .times. 240p 370.4K 25.63 26.35
36.60 43.09 43.13
[0259]
19TABLE 7 PSNR PSNY PSNR Bistream (PIC) Dimension Bit rate I-Qp
P-Qp -Y -Cb -Cr bicy_1000K_30F_G2_D31_Db1_ep0 720 .times. 480i
1047.7K 43.67 43.48 23.63 33.69 32.69 bus_1000K_30F_G2_D31_Db1_ep0
720 .times. 480i 968.6K 41.00 40.32 25.76 37.25 35.51
cheer_1000K_30F_G2_D31_Db1_ep0 720 .times. 480i 1045.7K 45.00 45.24
22.33 30.10 29.25 child_384K_30F_G2_D31_Db1_ep0 352 .times. 288p
382.7K 34.80 34.88 31.56 33.58 33.22
flower_1000K_30F_G2_D31_Db1_ep0 720 .times. 480i 996.3K 41.67 41.65
23.04 32.14 30.93 foreman_384K_30F_G2_D31_Db1_e- p0 352 .times.
288p 389.3K 30.00 30.34 34.31 41.35 39.88
foundtain_1000K_30F_G2_D31_Db1_ep0 320 .times. 240p 1149.5K 35.67
37.75 26.42 43.13 42.77 mino_384K_30F_G2_D31_Db1_ep0 320 .times.
240p 341.7K 28.17 28.27 35.61 38.95 37.21
news2_384K_30F_G2_D31_Db1_ep0 320 .times. 240p 418.4K 29.71 30.86
34.72 39.35 37.83 papa_384K_30F_G2_D31_Db1_ep0 320 .times. 240p
396.1K 24.33 24.27 40.32 44.36 43.16 pop_1000K_30F_G2_D31_Db1_ep0
720 .times. 480i 1174.5K 40.33 40.51 27.77 32.19 31.81
sasam_384K_30F_G2_D31_Db1_ep0 320 .times. 240p 372.1K 27.44 27.55
37.02 43.13 43.09
[0260]
20TABLE 8 Frame PIC Rate Control MB Rate Control No. Type: PIC-QP
BITS PSNR-Y Avg-QP BITS PSNR-Y 0 I: 35 72064 29.76 31.8 86728 30.75
1 P: 35 48128 28.26 34.9 39224 26.94 2 P: 35 54808 27.95 34.5 50200
27.24 3 P: 35 52120 28.04 34.7 43592 26.77 4 P: 35 54504 27.94 34.8
44464 26.53 5 P: 35 49088 28.18 34.8 41696 26.91 6 P: 35 48160 28.3
34.7 40232 26.95 7 P: 35 47560 28.37 34.7 40168 27.01 8 P: 35 48304
28.26 34.8 40336 26.84 9 P: 35 54048 27.97 34.8 45496 26.54 10 P:
35 48448 28.32 34.8 40104 26.88 11 P: 35 51904 28.06 34.8 42776
26.72 12 P: 35 50440 28.14 34.8 41512 26.68 13 P: 35 51600 28.07
34.7 43336 26.73 14 P: 35 50520 28.09 34.7 43264 26.77 15 P: 35
55576 27.99 34.7 47384 26.65 16 P: 35 54600 27.99 34.7 46464 26.62
17 P: 35 54080 28.03 34.8 44632 26.55 18 P: 35 51080 28.27 34.6
45528 27.38 19 P: 35 55472 28.06 34.4 49216 27.18 20 P: 35 54216
28.11 34.7 45600 26.65 21 P: 35 48784 28.37 34.8 40344 26.87 22 P:
35 54816 28.04 34.7 46272 26.62 23 P: 35 53664 28.17 34.7 44336
26.72 24 P: 35 50288 28.24 34.7 41840 26.79 25 P: 35 51480 28.18
34.6 43432 26.82 26 P: 35 49640 28.32 34.8 39184 26.83 27 P: 35
52608 28.12 34.6 44552 26.79 28 P: 35 53488 28.03 34.7 43968 26.59
29 P: 35 52712 28.07 34.7 43584 26.6 30 P: 35 48344 28.37 34.8
39024 26.89 31 P: 35 52712 28.09 34.7 43872 26.66 32 P: 35 54064
28.06 34.7 45344 26.61 33 P: 35 55904 27.95 34.7 46824 26.49 34 P:
35 55456 27.95 34.7 45744 26.52 35 P: 35 52984 28.17 34.6 44720
26.84 36 P: 35 55648 28 34.7 47792 26.66 37 P: 35 53176 28.13 34.8
43496 26.71 38 P: 35 48792 28.29 34.6 44072 27.54 39 P: 35 50712
28.2 34.4 45744 27.46 40 P: 35 49800 28.23 34.7 40728 26.84 41 P:
35 48720 28.23 34.8 40656 26.87 42 P: 35 52520 28.13 34.9 43376
26.63 43 P: 35 51504 28.16 34.6 46296 27.33 44 P: 35 52064 28.06
34.3 48688 27.32 45 P: 35 54720 27.99 34.8 45368 26.51 46 P: 35
56408 27.9 34.8 46288 26.44 47 P: 35 53488 28.09 34.7 43672 26.59
48 P: 35 51472 28.17 34.9 41136 26.64 49 P: 35 49984 28.2 34.5
45064 27.35 50 P: 35 57336 27.87 34.4 51760 27.08 51 P: 35 52688
28.04 34.8 43128 26.58 52 P: 35 55448 27.91 34.4 51328 27.16 53 P:
35 52560 28.05 34.7 44488 26.75 54 P: 35 50528 28.15 34.8 40880
26.71 55 P: 35 50672 28.11 34.8 42256 26.69 56 P: 35 51600 28.03
34.8 43112 26.68 57 P: 35 49392 28.27 34.7 40600 26.91 58 P: 35
50904 28.1 34.7 42208 26.73 59 P: 35 49520 28.25 34.7 41776 26.87
60 I: 35 69640 29.95 34.2 67336 29.36 61 P: 35 47624 28.43 34.7
38512 27.03 62 P: 35 55176 28.08 34.7 45880 26.66 63 P: 35 54592
27.94 34.8 45736 26.51 64 P: 35 54648 28.06 34.7 45520 26.62 65 P:
35 47424 28.31 34.7 40544 26.98 66 P: 35 51808 28.13 34.7 43704
26.73 67 P: 35 48824 28.34 34.7 40432 26.92 68 P: 35 47304 28.37
34.7 39488 27.04 69 P: 35 50104 28.17 34.7 41720 26.84 70 P: 35
52432 28.06 34.6 44232 26.79 71 P: 35 51136 28.16 34.6 42792 26.86
72 P: 35 47264 28.37 34.7 38992 27.04 73 P: 35 53824 27.99 34.7
45016 26.64 74 P: 35 48848 28.21 34.7 40096 26.85 75 P: 35 48584
28.17 34.7 40512 26.84 76 P: 35 46728 28.32 34.6 38592 27.01 77 P:
35 44840 28.35 34.8 37120 26.97 78 P: 35 48944 28.23 34.6 40856
26.89 79 P: 35 53432 28.03 34.6 44832 26.69 80 P: 35 48312 28.26
34.7 39584 26.87 81 P: 35 46616 28.4 34.7 38376 27.01 82 P: 35
50600 28.12 34.6 43272 26.89 83 P: 35 50024 28.21 34.7 40512 26.85
84 P: 35 47392 28.5 34.7 39512 27.11 85 P: 35 46360 28.49 34.7
38720 27.05 86 P: 35 48352 28.36 34.6 41544 27.02 87 P: 35 50528
28.24 34.6 42424 26.89 88 P: 35 46136 28.6 34.7 38520 27.22 89 P:
35 49152 28.26 34.7 41144 26.85 90 P: 35 46312 28.33 34.7 39008
27.04 91 P: 35 49544 28.2 34.6 41520 26.86 92 P: 35 47512 28.33
34.8 39152 27.02 93 P: 35 48984 28.24 34.4 45104 27.56 94 P: 35
44432 28.35 34.8 36384 27.07 95 P: 35 44448 28.39 34.8 36064 27.05
96 P: 35 49808 28.12 34.5 45320 27.37 97 P: 35 45480 28.29 34.9
37072 26.92 98 P: 35 45288 28.31 34.3 42568 27.65 99 P: 35 44216
28.41 34.8 36320 27.07 100 P: 35 45608 28.39 34.7 37176 27.07 101
P: 35 42000 28.53 34.8 34632 27.26 102 P: 35 43696 28.47 34.7 36056
27.2 103 P: 35 43576 28.48 34.8 35744 27.19 104 P: 35 46728 28.26
34.6 40728 27.04 105 P: 35 51568 28.13 34.8 42872 26.77 106 P: 35
48768 28.26 34.6 40656 26.96 107 P: 35 45248 28.39 34.8 36872 27
108 P: 35 50632 28.12 34.7 41840 26.76 109 P: 35 49232 28.2 34.7
41168 26.79 110 P: 35 44400 28.44 34.8 36200 27.07 111 P: 35 50552
28.05 34.6 43536 26.86 112 P: 35 47656 28.25 34.8 38744 27.01 113
P: 35 49376 28.23 34.5 42288 26.97 114 P: 35 50128 28.17 34.7 41208
26.82 115 P: 35 44864 28.43 34.7 36608 27.07 116 P: 35 44624 28.37
34.7 37024 27.02 117 P: 35 50968 28.19 34.7 41720 26.76 118 P: 35
49560 28.23 34.6 40960 26.93 119 P: 35 51400 28.11 34.7 42248 26.83
120 I: 35 68456 30.07 34.3 65816 29.38 121 P: 35 42520 28.66 34.7
34600 27.38 122 P: 35 45120 28.47 34.7 37208 27.12 123 P: 35 47280
28.34 34.6 39600 27.01 124 P: 35 47888 28.36 34.7 39208 26.95 125
P: 35 47648 28.43 34.8 38336 27 126 P: 35 43072 28.65 34.6 35672
27.3 127 P: 35 47944 28.35 34.6 40112 27.01 128 P: 35 45344 28.6
34.8 36200 27.12 129 P: 35 48384 28.37 34.5 41016 27.08 130 P: 35
45896 28.54 34.9 37816 27.09 131 P: 35 46952 28.39 34.8 38680 26.93
132 P: 35 43392 28.58 34.8 36432 27.2 133 P: 35 46656 28.35 34.6
38864 27.06 134 P: 35 48008 28.37 34.6 39216 26.97 135 P: 35 45600
28.46 34.8 36720 27.03 136 P: 35 52144 28.19 34.8 41680 26.75 137
P: 35 46648 28.44 34.8 37720 27.04 138 P: 35 46080 28.47 34.5 41136
27.67 139 P: 35 47568 28.36 34.3 42888 27.63 140 P: 35 50472 28.29
34.7 41416 26.79 141 P: 35 48800 28.31 34.6 41120 26.91 142 P: 35
53080 28.05 34.6 44920 26.71 143 P: 35 48016 28.41 34.9 40360 26.85
144 P: 35 47304 28.33 34.3 43240 27.62 145 P: 35 46072 28.44 34.9
37848 27.09 146 P: 35 47104 28.37 34.4 43440 27.71 147 P: 35 52144
28.1 34.6 43448 26.86 148 P: 35 52976 28.13 34.9 42920 26.64 149 P:
35 50848 28.22 34.5 45952 27.4 35 50107 28.26 34.66 42365 26.99
* * * * *